IRJET-Comparison of SIFT & SURF Corner Detector as Features and other Machine Learning Techniques for Identification of Commonly used Leaves

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 05 Issue: 03 | Mar-2018 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 6.171 | ISO 9001:2008 Certified Journal | Page 387
Comparison of SIFT & SURF Corner Detector as Features and other
Machine Learning Techniques for Identification of Commonly used
Leaves
Dr. Jharna Majumdar1, Anand Mahato2
1Dean R&D, Prof. & Head, Dept. of M.Tech CSE, Nitte Meenakshi Institute of Technology
Yelahanka, Bangalore-560 064, India
2 Student, M.Tech Sem IV, Department of M.Tech CSE, Nitte Meenakshi Institute of Technology
Yelahanka, Bangalore-560 064, India
----------------------------------------------------------------------------***----------------------------------------------------------------------------
Abstract— Scope of properly identifying the leaves very
vast. The properly identifying medicinal leaves for various
cures, identifying poisonous plants using leaves, determining
the usage of the plant using detected leaves are some of the
possible usages of leaf identification. The aim is to build a
methodology using various feature extraction techniques to
extract features, clustering algorithm to cluster the features
and decision trees as a classifier. All these methodologies put
together to form an effective method to efficiently recognize
the unknown leaf image using trained model. Feature
extraction techniques like SIFT and SURF which are robust
and provide matching in spite of the change in intensity, size
or rotation of the object in the images. Effective corner
points are chosen from the image from which magnitude
and orientation of surrounding are used to build descriptor
that is the vector of feature for each corner points. For
Clustering the data, various partitional, hierarchical, density
based methods are used to cluster the data which cluster the
data with respect to inter-connectivity, similarity, closeness,
etc. The clusters data is used to build the decision tree like
C4.5 and CART which uses entropy and Gini index as the
splitting criteria. The unknown image features are used to
traverse the decision tree of the closest cluster to yield the
matching image output from the training set.
Keywords—SIFT, SURF, ORB, Chameleon Clustering,
Decision Tree Classifier.
I. INTRODUCTION
Various types of plants and trees having different types
of leaves with a different shape, sizes, patterns, textures, etc
play an various different role in human life. The various
plants, trees or leaves can be identified effectively using
their respective leaf. Due to a millions variety of leaves and
plants available in the universe, it is not possible for
everyone to know each and every one. Since a lot of them
are very useful and many are useless, it is very useful to
correctly identify the plants using their leaves. Hence this
paper provides a mechanism for correctly training and
identifying commonly used leaves.
Object detection is the process of finding instances of
real-world objects such as faces, vehicles, intruder, etc in
images or videos. Object detection algorithms typically use
extracted features and learning algorithms to recognize
category or label of the object. It is commonly used in
applications such as image retrieval, security, surveillance,
and automated vehicle parking systems.
For feature extraction, various appearance-based and
feature based methods have been developed to overcome
various factors like illuminance, rotation, noise, scaling,etc.
Appearance based methods are relatively older, less
accurate and susceptible to various factors. It includes color
based methods, Edge matching, gradient matching,
geometrical shapes matching, etc. But evolving feature
based methods are very accurate, fast and robust. Methods
like SIFT, SURF, FAST, BRIEF, ORB, etc are the methods
which are most popular and vastly used nowadays for
object detection.
Clustering is one of the important aspect vastly used in
data mining and Machine Learning. Over the years, the
various algorithms have been evolved for clustering.
Partitioning based algorithms like K-means, PAM, CLARA,
etc., Hierarchical based methods like BIRCH, ROCK, CURE,
etc., Density based methods like DBSCAN are commonly
used popular clustering methods. In this paper, CURE
clustering has been used to cluster the features since CURE
is capable of handling large dataset, is relatively faster and
can find the clusters of various shapes and sizes.
Decision Trees is the one the mostly used classifier in
the Machine learning. There are various types of decision
trees like ID3, C4.5, CART, etc which are used for various
different purpose and data which are continuously
evolving. The decision trees are different due to splitting
criteria like entropy, Gini index, etc, various pruning
methods and other capabilities.
In this paper, SIFT or SURF is used for feature extraction
of leaf images which are then clustered using CURE
clustering algorithm whose decision tree is built using C4.5
or CART methods. The new leaf image is identified by
traversing the existing decision tree of the cluster for which
query features are having least distance.

II. LITERATURE REVIEW
A. SIFT CORNER DETECTOR
David Lowe presented a new methodology to obtain
distinctive invariant features from images in 2002 called
Scale Invariant Feature Transform (SIFT) [5]. These
features are invariant to translations, rotations, and scaling
of the image, and provides robust matching across
perspective transformations, distortions, change in 3D
viewpoints, noise, and illumination variations. They are
well localized in both the spatial and frequency domains,
reducing the probability of disruption by occlusion, clutter,
or noise. Large numbers of features can be extracted from
typical images with efficient algorithms. In addition, the
features are highly distinctive, which allows a single feature
to be correctly matched with high probability against a
large database of features, providing a basis for object and
scene recognition. The cost of extracting these features is
minimized by taking a cascade filtering approach, in which
the more expensive operations are applied only at locations
that pass an initial test. Following are the major stages of
computation used to generate the set of image features:
a. Scale-space extrema detection
The first stage of computation searches over all scales and
image locations. It is implemented efficiently by using a
difference-of-Gaussian function to identify potential
interest points that are invariant to scale and orientation.
Fig 1 depicts the Images in a different scales, and
compotation of DoG.
Fig.1 Images in a different scales, and compotation of DoG from [5]
b. Accurate Keypoint localization
At each candidate location, a detailed model is fit to
determine location and scale. Keypoints are selected based
on measures of their stability.
Once a keypoint candidate has been found by comparing a
pixel to its neighbors, the next step is to perform a detailed
fit to the nearby data for location, scale, and the ratio of
principal curvatures. This information allows points to be
rejected that have low contrast (and are therefore sensitive
to noise) or are poorly localized along an edge.
c. Orientation assignment
One or more orientations are assigned to each keypoint
location based on local image gradient directions to achieve
invariance to image rotation. A neighborhood is taken
around the keypoint location depending on the scale, and
the gradient magnitude and direction are calculated in that
region. An orientation histogram with 36 bins covering 360
degrees is created. (It is weighted by gradient magnitude
and Gaussian-weighted circular window with sigma equal
to 1.5 times the scale of keypoint. The highest peak in the
histogram is taken and any peak above 80% of it is also
considered to calculate the orientation. It creates keypoints
with same location and scale, but different directions. It
contributes to the stability of matching.
d. Keypoint descriptor
Here, keypoint descriptor for each keypoint is created. A
16x16 neighborhood around the keypoint is taken which is
divided into 16 sub-blocks of 4x4 size. For each sub-block, 8
bin orientation histograms are created. So, a total of 128 bin
values are available. It is represented as a vector to form
keypoint descriptor. In addition, several methods are
applied to achieve robustness against illumination changes,
rotation etc.
B. SURF CORNER DETECTORS
Speeded up robust features(SURF) algorithm [3] is the
feature point extraction algorithm purposed by Bay
H,TuytelaarsT,Gool L V in 2006. This algorithm is similar to
SIFT algorithm. SURF is a scale and rotation-invariant
interest point detector and a descriptor which is
computationally very fast. It uses Integral images to
improve the speed. The key points are detected by using a
Fast-Hessian matrix. The descriptor describes a
distribution of Haar-wavelet responses within the interest
point neighborhood. The SURF detector algorithm can be
summarized by following steps:
a. Fast-Hessian detector (interest point detection)
The performance in SURF can be ascribed to the use of an
intermediate image representation is known as the Integral
Image. The integral image is computed rapidly from an
input image and is used to speed up the calculation of any
upright rectangular area.
The SURF detector is based on the determinant of the
Hessian matrix. The determinant responses are normalized
to scale at this stage of the algorithm. The higher scale is
used the more pixels pass into the kernel and the largest
determinant response is received. This avoids the
probability that many features will be found in higher scale

when non-maximal suppression is used. The non-maximal
suppression, calculated at the fourth stage, is based on
finding the maximum determinant value within 26 nearest
neighbors in lower, present and upper scale. Afterwards,
this value is filtered with a predefined threshold to reserve
strongest interest point only.
b. Descriptor orientation assignment
At this stage, the Haar wavelets of size 4σ in x and y
directions are computed for the orientation assignment.
Wavelets are calculated for pixels that are located within a
radius of 6σ around the interest points. The dominant
orientation is evaluated by the sum of vertical and
horizontal responses. The descriptor is calculated at the last
stage, using Haar wavelets in square 20σ size area centered
at the interest point and oriented along the dominant
direction.
C. CURE CLUSTERING
(Clustering Using Representatives) CURE [1] is a bottom-up
hierarchical clustering algorithm, but instead of using a
centroid-based approach or an all-points approach it
employs a method that is based on choosing a well-formed
group of points to identify the distance between clusters.
Fig.2 shows the flow of the CURE clustering Algorithm.
Fig. 2 CURE Architecture
Instead of using single centroid or all points to represent a
cluster, a fixed number of representative points in a cluster
are chosen. The representative points are generated by
randomly selecting the scattered points from all points in
the cluster, then shrinking these scattered points toward
the mean of the cluster. CURE overcomes the problem of
favoring clusters with a spherical shape and similar sizes
and is more robust with respect to outliers.
The major steps of the CURE algorithm are as follows:
1. Pick up random sample of points from dataset
2. Cluster sample points hierarchically to create the initial
clusters
3. Pick representative Points
oFor each cluster pick k - representative point
dispersed across clusters
oMove each representative point a fixed fraction
towards the centroid of cluster
4. Labeling data
oRescan the whole dataset and visit each point in
dataset and place in the closest cluster
oThe closest cluster to the point p is determined by
comparing representative point closest to the point p.
The advantages of CURE are as follows:
• CURE can discover clusters with interesting shapes
•CURE is not very sensitive to outliers and the algorithm
filters them well
•Sampling and partitioning reduce input size without
sacrificing cluster quality
• Execution time of CURE is low compared to most of the
existing clustering algorithms
D.C4.5 DECISION TREE
Decision trees [16] are a very effective method of
supervised learning. It aims is the partition of a dataset into
groups as homogeneous as possible in terms of the variable
to be predicted.
C4.5 [9] is a decision tree technique for inducing
classification rules in the form of decision trees from a set
of given instances. C4.5 is a software extension of the basic
ID3 algorithm[8] designed by Quinlan (1979, 1983). Being a
supervised learning algorithm, it requires an arrangement
of preparing cases and every illustration can be viewed as a
pair: input object and the desired output value (class). The
classifier used by C4.5 is a decision tree and this tree is built
from root to leaves by choosing a simpler solution and a
certain splitting criterion.
Entropy and Information gain is the criteria which are used
in the C4.5 which is same for discrete variables as in ID3
algorithm but additionally, C4.5 supports continuous or
numeric data which splits node using a threshold of the
data in selected attribute yielding binary splits.
The major improvement of C4.5 decision tree over ID3 are:
(i) Accepts both continuous and discrete features.
(ii) Can handle missing data.

(iii) Tree pruning using various methods.
(iv) Different weights can be applied the features that
comprise the training data.
E. CART DECISION TREE
Classification And Regression Trees (CART) [17] is
developed by Bierman, Friedman, Olsten, Stone in early
80’s.CART is a nonparametric decision tree learning
technique can be used for building both Classification and
Regression Decision Trees. The impurity (or purity)
measure used in building a decision tree in CART is Gini
Index. The decision tree built by CART algorithm is always a
binary decision tree (each node will have only two child
nodes).
The main advantages are (i) CART handles missing values
automatically Using “surrogate splits” (ii) Invariant to
monotonic transformations of predictive variable (iii) Not
sensitive to outliers in predictive variables.
III. PROPOSED METHODOLOGY
Our Proposed method consists of the multi-step
methodology which consists of various feature extraction
techniques along with newer clustering method and
various types of decision trees for classification. Various
machine learning techniques and methodologieshave been
used to build arobust and effective framework for various
type object detection like commonly used leaves. The
proposed method also aims to do acomparative analysis of
SIFT and SURF feature extractions techniques along with
the C4.5 and CART decision tree for classification in our
given context.Proposed method is further divided into two
parts: Identification and Training.
A. TRAINING PHASE
As the name suggests, training of various types of
commonly used leaves dataset is used to train our machine
learning framework. Fig. 3 depicts the major steps involved
in training phase:
o Feature Extraction: For the interest point or features of
the training leaf images SIFT and SURF algorithms is
used once at a time and perform a comparative analysis
with respect to performance and accuracy.
o Clustering: Features extracted is merged to form a
training dataset to perform clustering as shown in
Figure 3. In this paper, CURE a hierarchical clustering
method is used to partition the training dataset into
specified k number of clusters.
o Decision Tree: At last, the decision tree is built for each
individual cluster formed. Here also two different
decision tree methods C4.5 and CART is used for relative
comparison between them.
Fig. 3 Flow diagram of Training Phase
B. IDENTIFICATION PHASE
After the completion of training, identification phase is used
to detect unknown leaf using our trained model .Fig 4.
Depicts the steps involved in Training Phase:
o Initially, the features are extracted from the query image
using SIFT or SURF
o Each feature keypoint is then compared with the
centroid of clusters (i.e mean of cluster points) formed
during training period using Euclidian distance
o Then, the respective decision tree of cluster matched is
traversed to find the matching label of that keypoint
o Finally, all the output from each keypoint is combined
and then voting is used to determine the matching leaf.

Fig. 4 Flow diagram of Testing Phase
IV. EXPERIMENT WORK, COMPARISON AND ANALYSIS
In the current work, set of 10, 20, 30 and 40 medical Leafs
with a different structure, shape, and sizes are taken as a
training dataset for given model. Also, the number of query
images 5,8,12 and 15 respectively for given dataset is
chosen from training images which are scaled and rotated.
The accuracy of the proposed work is calculated using a
number of query image correctly identified and percentage
of correct matching of the keypoints.
In addition, for the purpose of comparison, the multiple
numbers of experiments are carried out varying SIFT or
SURF for feature extraction with reduced keypoints along
with different decision trees i.e. C4.5 and CART used for
classification. Fig 5 shows the training leaf dataset.
Fig. 5 Training Leaf Dataset
The results of the experiments carried out are as follows:
Fig. 6 Comparison of Keypoint Accuracy of SIFT and SURF
From the number of experiments and observations the
performance of SIFT and SURF features along with decision
tree with entropy as splitting criteria has been found better
compared to others with respect to matching accuracy as
shown in the graph above. Although the percentage of
keypoints been matched for a selected sample leaf image
was found almost same for all others.
Fig.7 Matching Accuracy of SIFT and SURF
From the number of experiments and observations the
performance of SIFT and SURF features along with decision
tree with entropy as splitting criteria has been found better
compared to others with respect to matching accuracy as
shown in the graph above. Although the percentage of
keypoints been matched for a selected sample leaf image
was found almost same for all others.
V. CONCLUSION AND FUTURE WORK
Using the proposed work in this paper we have successfully
compared SIFT and SURF feature descriptors along with
Entropy and Gini index as decision tree criteria features for
the leaf image matching in given scenario. As per current
comparative analysis using proposed method, SIFT for the
feature extraction and C4.5 has been found suitable for the
best possible number of matches. The CURE clustering
algorithm has been successfully able to cluster feature data
with various dimensions.
The use of current proposed method gives us efficient
mechanism to match leaf images in training dataset. But

there is various new feature extraction techniques other
than SIFT and SURF that can perform better, various other
clustering techniques and decision trees yet to be tested in
our proposed method.
ACKNOWLEDGMENTS
Our sincere thanks goes to the Vision Group of Science and
Technology (VGST), Karnataka to acknowledge our
research and provide us the financial support to carry out
the research at NMIT. The authors express their sincere
gratitude to Ms. Shilpa Ankalaki for helping author during
preparation of the manuscript of the paper.
Finally our sincere gratitude goes to Prof. N R Shetty,
Director NMIT and Dr. H C Nagaraj, Principal NMIT for
providing the infrastructure support and whole hearted
encouragement to carry out the research at NMIT.
REFERENCE:
1. Guha, Sudipto, Rajeev Rastogi, and Kyuseok Shim.
"CURE: an efficient clustering algorithm for large
databases." ACM Sigmod Record. Vol. 27. No. 2.
ACM, 1998.
2. Xu, Rui, and Donald Wunsch. "Survey of clustering
algorithms." IEEE Transactions on neural
networks 16.3 (2005): 645-678.
3. Bay, Herbert, Tinne Tuytelaars, and Luc Van Gool.
"Surf: Speeded up robust features." Computer
vision–ECCV 2006 (2006): 404-417
4. Bay, Herbert, et al. "Speeded-up robust features
(SURF)." Computer vision and image
understanding 110.3 (2008): 346-359.
5. Lowe, D.: Distinctive image features from scale-
invariant keypoints, cascade filtering approach.
IJCV 60, 91–110 (2004)
6. He, Wei, et al. "Surf tracking." Computer Vision,
2009 IEEE 12th International Conference on. IEEE,
2009.
7. Safavian, S. Rasoul, and David Landgrebe. "A
survey of decision tree classifier methodology."
IEEE transactions on systems, man, and
cybernetics21.3 (1991): 660-674.
8. Quinlan, J. Ross. "Induction of decision trees."
Machine learning 1.1 (1986): 81-106.
9. Quinlan, J. Ross. C4. 5: programs for machine
learning. Elsevier, 2014.
10. Quinlan, J. Ross. "Improved use of continuous
attributes in C4. 5." Journal of artificial intelligence
research 4 (1996): 77-90.
11. Quinlan, J. Ross. "Bagging, boosting, and C4. 5."
AAAI/IAAI, Vol. 1. 1996.
12. Kohavi, Ronny, and J. Ross Quinlan. "Data mining
tasks and methods: Classification: decision-tree
discovery." Handbook of data mining and
knowledge discovery. Oxford University Press, Inc.,
2002.
13. Hui, Dong, and Han Dian Yuan. "Research of image
matching algorithm based on SURF fea-tures."
Computer Science and Information Processing
(CSIP), 2012 International Conference on. IEEE,
2012.
14. Hssina, Badr, et al. "A comparative study of
decision tree ID3 and C4. 5." International Journal
of Advanced Computer Science and Applications
4.2 (2014): 13-19.
15. Breiman, Leo. "Bagging predictors." Machine
learning 24.2 (1996): 123-140.
16. Bradski, Gary, and Adrian Kaehler. Learning
OpenCV: Computer vision with the OpenCV library.
" O'Reilly Media, Inc.", 2008

IRJET-Comparison of SIFT & SURF Corner Detector as Features and other Machine Learning Techniques for Identification of Commonly used Leaves

Recommended

Recommended

More Related Content

What's hot

What's hot (19)

Similar to IRJET-Comparison of SIFT & SURF Corner Detector as Features and other Machine Learning Techniques for Identification of Commonly used Leaves

Similar to IRJET-Comparison of SIFT & SURF Corner Detector as Features and other Machine Learning Techniques for Identification of Commonly used Leaves (20)

More from IRJET Journal

More from IRJET Journal (20)

Recently uploaded

Recently uploaded (20)

IRJET-Comparison of SIFT & SURF Corner Detector as Features and other Machine Learning Techniques for Identification of Commonly used Leaves